一种改进的线性区分分析方法及其在汉语数码语音识别上的应用

An Improved Linear Discriminant Analysis for Mandarin Digit Speech Recognition

尽管汉语数码语音识别只涉及十个数字 ,但由于不同数字的发音存在相同或相似的声母或韵母 ,造成汉语数码语音之间的混淆性很大 .采用通常的隐含马尔科夫模型 (HMM)作为汉语数码语音识别模型难以得到很高的识别率 .为了解决汉语数码之间的混淆问题 ,提高汉语数码语音识别性能 ,本文在隐含马尔科夫模型的状态层次上采用线性区分分析方法 ,将不同状态之间容易混淆的特征样本构成混淆模式类 ,针对混淆模式类进行线性区分分析 .通过线性区分变换 ,在变换特征空间中仅保留那些能够有效区分该混淆类别的特征参数 .这种基于状态的线性区分分析有效地提高了模型对混淆数码的区分能力 .实验表明即使采用状态数很少的粗糙识别模型 ,也能很大幅度提高模型的识别性能 ;经过线性区分变换优化后的汉语数码识别模型 ,孤立汉语数码语音识别率可以达到 99 32 % .

It is found that the phonetic similarities in the Mandarin digits are the main reasons for the difficulty of Mandarin digit recognition.In this paper,an improved linear discriminant analysis (LDA) based on the states of hidden Markov models (HMM) is presented.The recognition model discriminability is greatly improved by gathering the confusion data to the given states and then using the state-specific discriminative transformation.The experiments show that it increases the recognition rate greatly even if the simple models with insufficient states are used.The recognition accuracy of isolated Mandarin digits is over 99.32% after using optimal linear discriminative transformation.